EPS
← All tasks·#862Completed

artifacts: unified LoRA recipe + dose-to-target stopping (Phase 0e)

kind: infra
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Phase 0e of the unified artifact factory (plan: /home/thomasjiralerspong/.claude/plans/help-me-to-devise-vectorized-tower.md). Define the ONE unified LoRA recipe + its dose-to-target stopping mechanism that replaces the per-behavior recipe zoo (GENERIC / FACT / MARKER / WARMTH / turner_em) for all content/persona behaviors.

Goal

src/explore_persona_space/artifacts/recipe.py:

  • The single unified recipe — one TrainLoraConfig-override preset (proposed lr 1e-5, r32/α64, rsLoRA, contrastive negatives, generic-chat interleave) applied uniformly to content/persona behaviors.
  • Dose-to-target stopping — PRIMARY = checkpoint-and-select: train to a ceiling saving every K steps, then select the checkpoint whose SOURCE judged rate at C enters a preregistered mid-high band (~0.6–0.85, NOT the 1.0 ceiling — a ceiling censors the leakage read). OPTIONAL accelerator = an in-loop tf-margin proxy band-stop callback that MIRRORS eval/callbacks.py::MarkerBandStopCallback but reads eval/margin.compute_tf_margin (the #722-validated companion) to bound overshoot; confirm+select on the judged rate.
  • Programmatic carve-outs — marker keeps its marker-only-loss + log-prob band-stop [5,12] nat recipe; taught_fact keeps its span recipe. Route by the Behavior.programmatic flag / behavior name.
  • generic_frac knob (default a fixed modest fraction; 0 = the no-generic ablation) + the fullft matched-control path hook (train_method=fullft → ZeRO-3 via scripts/train_stage_sft.py, matched-dose).

Scope / constraints

  • Reuse train/sft.py::train_lora + TrainLoraConfig — the recipe is ONE config, do NOT reimplement training. The dose-to-target callback mirrors MarkerBandStopCallback (live in eval/callbacks.py).
  • This module DEFINES the recipe + stopping + carve-out routing; it does NOT drive a full training run (that's Phase 0g organisms.py).
  • Append the package export to artifacts/__init__.py append-only (sibling 0c/0f land concurrently).
  • WandB metrics required (code-style.md); NO dollar-budget caps (test_no_dollar_budget_caps.py).
  • CPU unit tests: recipe config builds; the dose-to-target callback fires on a synthetic rate/margin trajectory (mock); carve-out routing (marker→band-stop, fact→span recipe, content→unified). Lint + pytest.

Rules to read

.claude/rules/marker-training-recipe.md, .claude/rules/marker-leakage-measurement.md, .claude/rules/llm-judging.md (dose/DV), .claude/rules/on-policy-completions.md, .claude/rules/contrastive-negatives.md, .claude/rules/code-style.md.

Activity